AI news · Saturday, November 7, 2026

OpenAI's valuation hits $1.25 trillion amid AI jitters

OpenAI has raised $11 billion so far in a new funding round that values the company at $1.25 trillion, up from $852 billion in March. The round could grow to $20 billion, below the $30 billion OpenAI initially sought. The news lands right after tech stocks dropped nearly $2.5 trillion on AI-slowdown fears.

Follow the money: OpenAI scaled back its ask just as investors started questioning whether AI spending can keep growing this fast.

via The Verge

Voters just dealt data centers a major blow

In this week's US midterms, 41 local ballot measures asked voters about data center development. In all but one, the majority voted for more restrictions. Newly elected Democratic lawmakers are also expected to push for more scrutiny of AI companies and data centers, which require huge amounts of electricity and water.

What's next: This is the first time AI's physical footprint, not just its software, has become a mainstream voting issue nationwide.

via The Verge

Free Chinese AI model nearly matches top hacking tool

A new test had AI models try to break into mock company networks. OpenAI's best hacking model, available only to trusted partners, succeeded on 81% of tasks. China's free, downloadable DeepSeek model was close behind at 73%. Running DeepSeek for the task also cost $2 to $3, versus $17 to $46 for OpenAI's model.

The catch: Labs hoped restricting access to their best models would slow down hackers, but free alternatives are closing that gap fast.

via Wired

AI chatbot spending is flattening, even as usage climbs

New data from expense-management company Ramp shows AI spending barely grew, up just 1.5%, across August and September, even though usage kept rising. That's a sharp slowdown from July's 42% monthly growth. Price cuts from OpenAI and Anthropic, plus cheaper rivals, mean companies are using more AI for less money.

The twist: Rising usage but flat revenue suggests AI may be turning into a cutthroat, low-margin business rather than a hypergrowth one.

via TechCrunch

Google's new video AI makes 15-second clips for 30% less

Google's Gemini Omni 1.2 Flash can now generate 15-second videos, up from 10 seconds, and keeps faces, outfits and logos consistent across multiple camera angles. You can also edit a clip by just chatting with it instead of starting over. The price drops 30% from the previous version, to $0.07 per second of 720p video.

What's next: Cheaper, more consistent AI video is the kind of unglamorous upgrade that quietly makes AI-made ads and product videos viable at scale.

via Google AI

Companies are paying up to $2 million for your old work emails

A growing industry is buying companies' internal documents, emails and Slack messages to train AI models, with offers reportedly ranging from $350,000 to $2 million. Mercor, a leader in the space founded in 2023, had its three cofounders become billionaires last year. One firm even offered a freelance writer $1,000 for a single story.

Follow the money: As public internet text runs out, AI labs are shifting to buying the private, messy data of how real workplaces actually function.

via Wired

Meta drops controversial plan to scrape linked websites

Meta has paused a planned change that would have forced Facebook business pages to let Meta's AI crawl any external website they linked to, even bypassing paywalls. The plan faced lawsuits in Brazil and Norway, a UK antitrust investigation, and bipartisan pushback from US senators. Meta says existing terms stay in place while it consults publishers.

The catch: The backlash shows regulators and courts, not just public pressure, are now actively blocking how AI companies try to acquire training data.

via Ars Technica

A startup's humanoid robot cleaned a reporter's apartment

San Francisco startup Tau Robotics sent a four-and-a-half-foot humanoid robot to clean an MIT Technology Review writer's apartment for $30 an hour, now $60 for weekly visits with two robots. The robot is actually remote-controlled by a human, not autonomous, and struggled badly with stairs. Every task it completes trains Tau's future self-operating models.

The catch: What looks like a slick robot demo is really cheap human labor in a robot suit, collecting the data needed to eventually replace itself.

via MIT Technology Review

Open models get new tests for Egyptian and Nigerian accents

A speech-recognition leaderboard run by Hugging Face has added two new accent datasets, Egyptian and Nigerian English, built from real conversational speech. The project now sorts accents using a system that separates US, UK, Australian and Canadian English from accents like India's, Nigeria's and Egypt's, so researchers can see exactly where models fall short.

For you: Most voice AI is tuned for US and UK accents first, so this kind of testing is what eventually gets your own accent recognized properly.

via Hugging Face

What is a GPU, and why does every AI company need them?

A GPU, or graphics processing unit, is a computer chip built to do thousands of simple calculations at once, instead of one complex calculation at a time like a regular computer chip (a CPU). That was originally useful for drawing video game graphics, where millions of pixels need updating simultaneously.

For you: Nearly every AI headline about chip shortages, export bans, or data center costs is really a story about this one piece of hardware.

Original explainer

China allegedly recruited exec to spy on Nvidia's CEO

China's intelligence agency recruited Megaspeed executive Alice Huang to gather information about American AI chips and potentially influence Nvidia to sell its most advanced chips to China, according to the New York Times. Huang has close ties to Nvidia CEO Jensen Huang and other top chip executives. Three House committees have asked Huang to brief them on the relationship.

Zoom out: It's a sign of how high the stakes have gotten in the US-China fight over who controls the chips that power AI.

via The Verge

A 100-task test checks if AI can actually do lab science

OpenAI built BioBasicsBench, a set of 100 tasks that measure whether AI agents can interpret raw data from everyday biology lab experiments, like reading a gel or spotting a contaminated sample. It's meant to test real, basic lab reasoning rather than just answering science trivia, an early step toward AI that can genuinely assist researchers.

For you: Useful lab AI depends on handling messy, ambiguous real data, not just acing clean textbook questions, which is a much harder bar.

via OpenAI